Wai Yip Tung Wai Yip Tung - 4 months ago 595x
Python Question

PySpark add a column to a DataFrame from a TimeStampType column

I have a DataFrame that look something like that. I want to operate on the day of the


|-- host: string (nullable = true)
|-- user_id: string (nullable = true)
|-- date_time: timestamp (nullable = true)

I tried to add a column to extract the day. So far my attempts have failed.

df = df.withColumn("day", df.date_time.getField("day"))

org.apache.spark.sql.AnalysisException: GetField is not valid on fields of type TimestampType;

This has also failed

df = df.withColumn("day", df.select("date_time").map(lambda row: row.date_time.day))

AttributeError: 'PipelinedRDD' object has no attribute 'alias'

Any idea how this can be done?


You can use simple map:

df.rdd.map(lambda row: Row(**dict(row.asDict(), day=row.date_time.day)))

Another option is to register a function and run SQL query:

sqlContext.registerFunction("day", lambda x: x.day)
sqlContext.registerDataFrameAsTable(df, "df")
sqlContext.sql("SELECT *, day(date_time) as day FROM df")

Finally you can define udf like this:

from pyspark.sql.functions import udf
from pyspark.sql.types import IntegerType

day = udf(lambda date_time: date_time.day, IntegerType())
df.withColumn("day", day(df.date_time))


Actually if you use raw SQL day function is already defined (at least in Spark 1.4) so you can omit udf registration. It also provides a number of different date processing functions including:

It is also possible to use simple date expressions like:

current_timestamp() - expr("INTERVAL 1 HOUR")

It mean you can build relatively complex queries without passing data to Python. For example:

df =  sc.parallelize([
    (1, "2016-01-06 00:04:21"),
    (2, "2016-05-01 12:20:00"),
    (3, "2016-08-06 00:04:21")
]).toDF(["id", "ts_"])

now = lit("2016-06-01 00:00:00").cast("timestamp") 
five_months_ago = now - expr("INTERVAL 5 MONTHS")

    # Cast string to timestamp
    # For Spark 1.5 use cast("double").cast("timestamp")
    .withColumn("ts", unix_timestamp("ts_").cast("timestamp"))
    # Find all events in the last five months
    .where(col("ts").between(five_months_ago, now))
    # Find first Sunday after the event
    .withColumn("next_sunday", next_day(col("ts"), "Sun"))
    # Compute difference in days
    .withColumn("diff", datediff(col("ts"), col("next_sunday"))))